In the SL domain, NN-based approaches have proven effective in solving difficult problems. However, these networks, and particularly DNNs, require a large quantity of annotated images. In fact, many apps don't have all images labeled. Images must be simultaneously recognized and annotated as they go. Therefore, learning must be carried out continuously. The user supervises the entire training phase and corrects errors in order to gradually retrain the networks with the corrected data. In this new use, the objective to be achieved is not the lowest error rate, but the reduction of the total time to annotate, correct and train the neural networks. To evaluate the effectiveness of our approach, we will compare the total time to achieve recognition by an continual approach with a classic deep learning approach with the times of manual annotation of the entire ground truth and error correction .